Multilabel Strategy Playground

Design a small module that trains and serves multi-class and multi-label predictors using reduction strategies. Emphasis is on using built-in tooling from the declared dependency rather than hand-rolled loops.

Capabilities

Independent label prediction

• With training features [[0, 0], [1, 1], [1, 0], [0, 1]], label names ["primary", "secondary"], and binary targets [[0, 0], [1, 1], [1, 0], [0, 1]], fitting the independent-model trainer and predicting for [[1, 0], [0, 1]] should yield [["primary"], ["secondary"]] when using the default threshold of 0.5. @test
• With the same training data, predicting for [[1, 1]] using a threshold of 0.8 should return ["primary", "secondary"] as both labels clear the higher confidence cutoff. @test

Chained label prediction

• Using training features [[0], [1], [2], [3]] with label names ["base", "bonus"] and binary targets [[0, 0], [1, 0], [1, 1], [1, 1]], fitting the chain-based trainer with explicit order [0, 1] and predicting for [[0.5], [2.5]] should yield [[], ["base", "bonus"]]. @test

Pairwise multiclass voting

• With training features [[0], [1], [2], [3]] and targets [0, 0, 1, 2], fitting the pairwise multiclass reducer and predicting for [[0.2], [2.6]] should yield [0, 2]. @test

Implementation

@generates

• Use deterministic training (fixed random seed) so predictions are repeatable across runs.
• Choose base learners that expose calibrated probability or decision scores so thresholds meaningfully gate predicted labels.

API

from typing import Any, Dict, List, Optional, Sequence, Tuple

Label = str

def train_independent(
    X_train: Sequence[Sequence[float]],
    Y_train: Sequence[Sequence[int]],
    label_names: Sequence[Label]
) -> Any:
    """Fits and returns a multi-label model built from independent binary problems."""

def train_chained(
    X_train: Sequence[Sequence[float]],
    Y_train: Sequence[Sequence[int]],
    label_names: Sequence[Label],
    order: Optional[Sequence[int]] = None
) -> Any:
    """Fits and returns a dependency-aware chain model using the provided or inferred label order."""

def predict_labels(
    model: Any,
    X: Sequence[Sequence[float]],
    label_names: Sequence[Label],
    threshold: float = 0.5
) -> List[List[Label]]:
    """Predicts label sets for each sample using the provided fitted model."""

def train_pairwise(
    X_train: Sequence[Sequence[float]],
    y_train: Sequence[int]
) -> Any:
    """Fits and returns a multiclass model built from pairwise binary reductions."""

def predict_class(
    model: Any,
    X: Sequence[Sequence[float]]
) -> List[int]:
    """Predicts a single class label for each sample using the pairwise model."""

Dependencies { .dependencies }

scikit-learn { .dependency }

Provides multioutput reduction strategies and base estimators.